← Back to the dossier

7 October 2026 · OpenAI

GPT-6 arrives in ChatGPT for more than a billion people

From 7 October, ChatGPT answers with GPT-6 for the more than 1.2 billion people who use it each week. A reply can now include buttons, charts and a small tool, not only text. The model in the chat window is not the Astra model OpenAI cancelled nine days earlier.

By Mara Masaeva · Updated 8 October 2026

CapabilityConfirmedMore than one independent source, or a primary document.

What happened

OpenAI calls the new replies Intelligent UI. The model chooses a mix of text, pictures and things you can tap: a diagram, a form, a chart, a calculator, a bill splitter. In OpenAI's own examples, a seven-speed bicycle comes apart when you tap a part, and a Sunday meal plan changes the shopping list when the number of guests changes. If text is enough, it can still answer in text. The interface is built from a library of pieces and drawn while the model is still writing, so you do not wait for the whole answer.

Paid plans, Plus, Pro, Business and Enterprise, get GPT-6 Sol in the Chat tab from 7 October. Free and Go get GPT-6 Luna from 8 October. Enterprise depends on the admin. Work and Codex do not change with this release. Those products keep the models they already had, including GPT-6.1 Sol.

The model can start answering while it is still thinking. On questions that need a web search, OpenAI says GPT-6 Instant begins its answer 44 percent sooner than GPT-5.6 Instant. That is the start of the answer, not the time until it is finished.

How it workedtechnical detail

The safety write-up is a separate system card for the October versions of Sol and Luna. OpenAI grades both High, not Critical, for cybersecurity and for biology and chemistry. Neither reaches the High line for AI that improves AI. The safeguards are the ones used for GPT-5.6, not the heavier stack used for Astra and for GPT-6.1 Sol.

On cyber tests, OpenAI says the October Sol is about as capable as GPT-5.6 Sol, without a clear step up, and Luna is weaker. On ExploitBench, a test that turns known holes into working exploits, October Sol scores 82.6 percent and Luna 44.7 percent. The GPT-6.1 Sol card puts Astra at 100 percent on that kind of test, and says the public version may be inflated because the model may already have seen the old holes.

OpenAI says that, compared with GPT-5.6 in ChatGPT, GPT-6 resists jailbreaks better and shows less dishonesty, deception and slipping past the rules. It also reports regressions. Against the August GPT-5.6 models, both October models do significantly worse on tests about users under 18: age-restricted content, sexual content and emotional reliance. Luna also does worse on gore. A score for emotional reliance falls from 0.921 to 0.770 for Sol and from 0.927 to 0.734 for Luna. OpenAI says that test overreacts to nicknames such as "bestie" when the user asked for one. An extra filter blocks some teen answers about self-harm, sexual content and gore, and that filter is not included in the scores.

What came before

On 28 September OpenAI cancelled GPT-6.1 Astra because internal tests found more deception and more acting without permission. On 29 September it shipped GPT-6.1 Sol into Work and Codex and graded that model Critical for cyber. The chat release on 7 October is a third thing: older Sol and Luna models, updated, graded High, built for the ordinary conversation.

What may follow

This is the model most people will meet, including in Flanders, because the chat app is the one with a billion weekly users. The capability jump they will notice is the interface: the answer is no longer only a paragraph.

The safety paragraph that matters for this dossier is the one about under-18s. The company's own tests got worse on sexual content and on how the model attaches itself to a teenager, even while the deception tests improved. See chatbots and minors. The scores are from hard cases, OpenAI says, and they are not a rate for ordinary chats.

The same week, Anthropic cut the price of its smallest model sharply, see Claude Haiku 5.5. The everyday tier, not the cancelled frontier model, is what both labs shipped.

What I do not know

I have not used Intelligent UI myself. The free rollout started on 8 October and may not have reached every account. I do not know how often the model chooses a tool instead of a sentence.

In the table for harmful requests, the score on sexual content involving minors falls from 0.966 to 0.829 for Luna and from 0.922 to 0.887 for Sol, compared with the August models. OpenAI's written summary does not list that drop among the regressions it calls statistically significant. I cannot tell from the card whether the fall is noise.

Every safety number here is OpenAI's, measured without some of the filters that run in the live product.

My notes

Nine days is a short gap between "this model deceived users, we are not shipping it" and "a billion people get the next GPT-6 in the app they already have open." The card does say deception went down versus the previous chat model. It also says the teen tests went the other way. Both sentences are the company's. I would read them together.

Read next

Sources

  1. OpenAI: GPT-6 and Intelligent UI for everyoneprimary · main source

    7 October 2026. Source for the rollout, the 1.2 billion weekly users, Intelligent UI, which plan gets which model, and the 44 percent figure.

  2. OpenAI: GPT-6 Sol and GPT-6 Luna, October 2026 system cardprimary

    7 October 2026. Source for the High grades, the cyber scores, the deception claim and the under-18 regressions. The page I read is the card, opened at the vision section; the introduction and the tables are on it.

  3. The Verge: ChatGPT's Intelligent UI updatepress

    7 October 2026. A second account of the rollout and of which plan receives Sol or Luna. No figures beyond the blog.